YOLOv2: represents the current level of the industry's most advanced object detection, other detection systems (FasterR-CNN, ResNet, SSD) it's faster too, users can trade-off between its speed and accuracy.
YOLO9000: This network structure can be detected in real time more than 9000 kinds of object classification, thanks to its use WordTree, by mixing WordTree to detect data within the data set to identify a data set.
The current test data set (Detection Datasets) information classification label is too small, the picture is less than the number of classified data sets, cost test data set is too high, it can not be classified as a data set to use. And now classified data set but has a lot of pictures and very rich category.
We proposed a new method of training - joint training algorithm. Such an algorithm may be mixed together these two data sets. Using a hierarchical view of the object of classification, the classified data set with a huge amount of data sets to expand the detection data, whereby the mixed two different data sets.
The basic idea of joint training algorithm is this: while training in the detection data sets and classified data set for the object detector (Object Detectors), the exact location of a detected data set of data for learning objects, data classification data sets to increase the classification of volume, improve robustness.
YOLO9000 is the use of joint training algorithm trained, he has a kind of classification information 9000, which classified information is learned from ImageNet classification datasets, while learning from the object position detection COCO test data sets.